ATI Radeon HD 4670 vs NVIDIA GeForce 9600 GSO

The GeForce 9600 GSO GPU has better pixel fill rate due to greater number of Raster Operations Pipelines (ROPs), even though its graphics frequency is lower. Having higher pixel fillrate allows the card to draw more pixels on screen and off screen, which is beneficial for some 3D effects in games, or when playing at higher display resolutions.

Although it has slower graphics clock, the GeForce 9600 GSO GPU comes with many more TMUs, resulting in better texture fill rate. Better maximum texture fill rate means that the card can utilize more complex 3D effects and/or map more textures to each texel, which improves visual appearance of games and generated images.

Single Precision performance is useful for estimating card's maximum speed in programs, that process only single-precision floating point data. This performance is expressed in GFLOPS, or Giga (billions of) Floating Point Operations Per Second. All single-precision calculations are performed by stream processors or CUDA cores, as such the more cores / processors the graphics unit has the better. Performance also depends on the processor clock speed. The Radeon HD 4670 is substantially faster here. Higher single-precision performance number means the GPU will perform better in general computing applications. Since stream processors or CUDA cores are also used as vertex and geometry shaders for 3D image generation, higher performance is also beneficial to games.

To speed up processing, the graphics cards store 3D scene data, textures and intermediate data, used for image generation, in on-board memory. The video memory usually has much higher bandwidth than system RAM, and more bandwidth allows the graphics card to run at higher screen resolutions, use larger and more detailed textures, and apply more complex 3D effects and filters. The bandwidth is dependent on several components, like memory type, speed, and memory interface width. Specifically, higher memory bandwidth of the GeForce 9600 GSO graphics card is a direct result of wider memory bus.